Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› restud-tables-figures
restud-tables-figures
GitHub将《经济研究评论》手稿中的表格和图表优化为出版级标准。确保符合期刊优雅简洁风格，使用向量化图形、专业排版及自包含注释，并通过代码自动生成以确保可复现性。
触发场景
表格列数过多或脚注冗长导致阅读困难
核心结果被淹没在表格中，更适合用图表展示
图表为光栅格式或缺乏自包含说明
各图表间符号、单位或样本量不一致
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill restud-tables-figures -g -y
SKILL.md
Frontmatter
{
"name": "restud-tables-figures",
"description": "Use when the exhibits in a The Review of Economic Studies (REStud) manuscript need to be made publication-grade — regression tables that are oversized or footnote-bloated, or figures that do not carry the result. Finalizes exhibits to REStud house standard; does not run the underlying analysis."
}
REStud Tables & Figures (restud-tables-figures)
When to trigger
- A main table has more columns than a reader can hold in mind (≳ 6)
- Tables carry default software output (stars only, no economic-magnitude reading)
- The central result is buried in a table when a figure would carry it
- Figures are raster (PNG) rather than vector, or lack a self-contained notes block
- Notation, units, or sample counts are inconsistent across exhibits
REStud exhibit principles
REStud values elegance and economy, and that extends to exhibits. Two REStud-specific facts: (1) auxiliary exhibits go in the online appendix, so the main-text figure budget is tight — promote only the exhibits that carry the contribution; (2) every published exhibit must map to a line in the deposited code, because the Data Editor regenerates tables and figures from your replication package before publication (AEA DCAS check; see restud-replication-package) — so script your exhibits, never hand-edit them. The standard:
- Each exhibit carries one result. If a table is doing two jobs, split it.
- The reader can read the magnitude, not just the stars. State units; an effect of "0.034***" is meaningless until the reader knows 0.034 of what.
- Figures are first-class. A striking new fact is often best shown as a figure (event-study plot, binscatter, RD plot, model-vs-data overlay). For a new-fact paper, the headline figure may matter more than any table.
- Self-contained notes. Every table and figure has a notes block stating sample, period, unit of observation, fixed effects, standard-error clustering, and significance convention — readable without the main text.
Tables
- Use professional rules (booktabs
\toprule/\midrule/\bottomrule); no vertical rules, no double horizontal rules. - Report coefficient, standard error (in parentheses), and N; add the dependent-variable mean so magnitudes are interpretable.
- Group columns by specification with clear panel headers; indicate fixed effects with a Yes/No block at the foot, not buried in notes.
- Significance: report SEs and let the reader judge; if stars are used, define them once and consistently.
- Generate from code (
estout/esttab,etable,modelsummary) so the table regenerates with the result — no hand-typed numbers.
Figures
- Vector format (PDF / EPS) — never raster for line art.
- Confidence bands on every event-study / RD / dose-response plot.
- Axis labels with units; legend inside the plot area or a one-line caption; no chartjunk, no 3-D, no gradient fills.
- For a new-fact paper, lead with the figure that is the fact.
- Color must survive grayscale printing; distinguish series by marker/line style as well as hue.
Theory exhibits
For theory and theory-with-empirics papers, exhibits earn their place too:
- A figure of the model's key mechanism (best-response curves, a phase diagram, the comparative-static frontier) can carry a proposition better than the algebra.
- A model-vs-data overlay is among the most persuasive REStud exhibits: plot the model's prediction against the empirical moments it was not fit to match.
- Numerical illustrations and calibrated simulations belong in a clearly labeled figure with the parameter values in the notes or the online appendix.
Consistency pass
Before finalizing, run one pass across all exhibits together (not one at a time):
- Variable names, units, and sample sizes match across every table and figure.
- The same coefficient reported in two places shows the same number to the same precision.
- Numbering is continuous and every exhibit is referenced in the text in order.
Inconsistent exhibits are a fragility signal — referees who spot a mismatched N start doubting the rest.
Execution bridge (StatsPAI / Stata MCP)
Generate exhibits from the fitted result, not by retyping numbers. Full map:
execution-with-mcp. REStud is top-5 general-interest economics; credible identification with modern estimators is the bar across applied fields.
- Tables:
etable(multi-model) ordid_summary_to_latexstraight from theresult_id. - Figures:
plot_from_result/enhanced_event_study_plot/event_study_table— axis units and the SE/clustering note baked in. - Every note names the estimator + clustering and states the magnitude in interpretable units.
See a full fitted-result → exhibit chain in the JF execution walkthrough.
Checklist
- Each exhibit carries exactly one result
- Main table ≤ ~6 columns; longer variants moved to the appendix
- Dependent-variable mean (or comparable scale) reported for magnitude reading
- FE / clustering / sample shown in a foot block or self-contained notes
- Figures are vector; event-study/RD plots show confidence bands
- Exhibits regenerate from code — no hand-typed numbers
- Notation, units, and N consistent across all exhibits
Anti-patterns
- A 14-column table with no headline takeaway
- Reporting only stars, so the economic magnitude is unreadable
- Raster figures that pixelate in the typeset PDF
- Notes that say "see text" instead of being self-contained
- Burying the paper's striking fact in a table when a figure would make it obvious
- Inconsistent variable names or sample sizes across tables (a fragility signal to referees)
Output format
【MAIN EXHIBITS】[table/figure — the one result each carries]
【MAGNITUDE READABLE】yes / no — scale reported
【FORMAT】booktabs tables / vector figures — confirmed
【NOTES SELF-CONTAINED】yes / no
【HEADLINE FIGURE】<which exhibit is the paper's "face">
【NEXT SKILL】restud-writing-style
版本历史
- 1839142 当前 2026-07-05 14:21
